OBJECTIVES: To evaluate haemoglobin A1c (HbA1c) in diagnosing diabetes and identify the optimal HbA1c threshold to be used in Chinese adults. DESIGN: Multistage stratified cross sectional epidemiological survey. SETTING: Shanghai, China, 2007–8. PARTICIPANTS: 4886 Chinese adults over 20 years of age with no history of diabetes. MAIN OUTCOME MEASURES: Performance of HbA1c at increasing thresholds for diagnosing diabetes. RESULTS: The area under the receiver operating characteristics curve for detecting undiagnosed diabetes was 0.856 (95% confidence interval 0.828 to 0.883) for HbA1c alone and 0.920 (0.900 to 0.941) for fasting plasma glucose alone. Very high specificity (96.1%, 95% confidence interval 95.5% to 96.7%) was achieved at an HbA1c threshold of 6.3% (2 SD above the normal mean). Moreover, the corresponding sensitivity was 62.8% (57.1% to 68.3%), which was equivalent to that of a fasting plasma glucose threshold of 7.0 mmol/l (57.5%, 51.7% to 63.1%) in detecting undiagnosed diabetes. In participants at high risk of diabetes, the HbA1c threshold of 6.3% showed significantly higher sensitivity (66.9%, 61.0% to 72.5%) than both fasting plasma glucose ≥7.0 mmol/l (54.4%, 48.3% to 60.4%) and HbA1c ≥6.5% (53.7%, 47.6% to 59.7%) (P<0.01). CONCLUSIONS: An HbA1c threshold of 6.3% was highly specific for detecting undiagnosed diabetes in Chinese adults and had sensitivity similar to that of using a fasting plasma glucose threshold of 7.0 mmol/l. This optimal HbA1c threshold may be suitable as a diagnostic criterion for diabetes in Chinese adults when fasting plasma glucose and oral glucose tolerance tests are not available.